mm/filemap.c

   1 /*
   2  *      linux/mm/filemap.c
   3  *
   4  * Copyright (C) 1994-1999  Linus Torvalds
   5  */
   6
   7 /*
   8  * This file handles the generic file mmap semantics used by
   9  * most "normal" filesystems (but you don't /have/ to use this:
  10  * the NFS filesystem used to do this differently, for example)
  11  */
  12 #include <linux/config.h>
  13 #include <linux/module.h>
  14 #include <linux/slab.h>
  15 #include <linux/compiler.h>
  16 #include <linux/fs.h>
  17 #include <linux/aio.h>
  18 #include <linux/kernel_stat.h>
  19 #include <linux/mm.h>
  20 #include <linux/swap.h>
  21 #include <linux/mman.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/file.h>
  24 #include <linux/uio.h>
  25 #include <linux/hash.h>
  26 #include <linux/writeback.h>
  27 #include <linux/pagevec.h>
  28 #include <linux/blkdev.h>
  29 #include <linux/security.h>
  30 #include <linux/syscalls.h>
  31 #include "filemap.h"
  32 /*
  33  * FIXME: remove all knowledge of the buffer layer from the core VM
  34  */
  35 #include <linux/buffer_head.h> /* for generic_osync_inode */
  36
  37 #include <asm/uaccess.h>
  38 #include <asm/mman.h>
  39
  40 /*
  41  * Shared mappings implemented 30.11.1994. It's not fully working yet,
  42  * though.
  43  *
  44  * Shared mappings now work. 15.8.1995  Bruno.
  45  *
  46  * finished 'unifying' the page and buffer cache and SMP-threaded the
  47  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
  48  *
  49  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
  50  */
  51
  52 /*
  53  * Lock ordering:
  54  *
  55  *  ->i_mmap_lock               (vmtruncate)
  56  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
  57  *      ->swap_list_lock
  58  *        ->swap_device_lock    (exclusive_swap_page, others)
  59  *          ->mapping->tree_lock
  60  *
  61  *  ->i_sem
  62  *    ->i_mmap_lock             (truncate->unmap_mapping_range)
  63  *
  64  *  ->mmap_sem
  65  *    ->i_mmap_lock
  66  *      ->page_table_lock       (various places, mainly in mmap.c)
  67  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
  68  *
  69  *  ->mmap_sem
  70  *    ->lock_page               (access_process_vm)
  71  *
  72  *  ->mmap_sem
  73  *    ->i_sem                   (msync)
  74  *
  75  *  ->i_sem
  76  *    ->i_alloc_sem             (various)
  77  *
  78  *  ->inode_lock
  79  *    ->sb_lock                 (fs/fs-writeback.c)
  80  *    ->mapping->tree_lock      (__sync_single_inode)
  81  *
  82  *  ->i_mmap_lock
  83  *    ->anon_vma.lock           (vma_adjust)
  84  *
  85  *  ->anon_vma.lock
  86  *    ->page_table_lock         (anon_vma_prepare and various)
  87  *
  88  *  ->page_table_lock
  89  *    ->swap_device_lock        (try_to_unmap_one)
  90  *    ->private_lock            (try_to_unmap_one)
  91  *    ->tree_lock               (try_to_unmap_one)
  92  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
  93  *    ->private_lock            (page_remove_rmap->set_page_dirty)
  94  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
  95  *    ->inode_lock              (page_remove_rmap->set_page_dirty)
  96  *    ->inode_lock              (zap_pte_range->set_page_dirty)
  97  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
  98  *
  99  *  ->task->proc_lock
 100  *    ->dcache_lock             (proc_pid_lookup)
 101  */
 102
 103 /*
 104  * Remove a page from the page cache and free it. Caller has to make
 105  * sure the page is locked and that nobody else uses it - or that usage
 106  * is safe.  The caller must hold a write_lock on the mapping's tree_lock.
 107  */
 108 void __remove_from_page_cache(struct page *page)
 109 {
 110         struct address_space *mapping = page->mapping;
 111
 112         radix_tree_delete(&mapping->page_tree, page->index);
 113         page->mapping = NULL;
 114         mapping->nrpages--;
 115         pagecache_acct(-1);
 116 }
 117
 118 void remove_from_page_cache(struct page *page)
 119 {
 120         struct address_space *mapping = page->mapping;
 121
 122         BUG_ON(!PageLocked(page));
 123
 124         write_lock_irq(&mapping->tree_lock);
 125         __remove_from_page_cache(page);
 126         write_unlock_irq(&mapping->tree_lock);
 127 }
 128
 129 static int sync_page(void *word)
 130 {
 131         struct address_space *mapping;
 132         struct page *page;
 133
 134         page = container_of((page_flags_t *)word, struct page, flags);
 135
 136         /*
 137          * page_mapping() is being called without PG_locked held.
 138          * Some knowledge of the state and use of the page is used to
 139          * reduce the requirements down to a memory barrier.
 140          * The danger here is of a stale page_mapping() return value
 141          * indicating a struct address_space different from the one it's
 142          * associated with when it is associated with one.
 143          * After smp_mb(), it's either the correct page_mapping() for
 144          * the page, or an old page_mapping() and the page's own
 145          * page_mapping() has gone NULL.
 146          * The ->sync_page() address_space operation must tolerate
 147          * page_mapping() going NULL. By an amazing coincidence,
 148          * this comes about because none of the users of the page
 149          * in the ->sync_page() methods make essential use of the
 150          * page_mapping(), merely passing the page down to the backing
 151          * device's unplug functions when it's non-NULL, which in turn
 152          * ignore it for all cases but swap, where only page->private is
 153          * of interest. When page_mapping() does go NULL, the entire
 154          * call stack gracefully ignores the page and returns.
 155          * -- wli
 156          */
 157         smp_mb();
 158         mapping = page_mapping(page);
 159         if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
 160                 mapping->a_ops->sync_page(page);
 161         io_schedule();
 162         return 0;
 163 }
 164
 165 /**
 166  * filemap_fdatawrite_range - start writeback against all of a mapping's
 167  * dirty pages that lie within the byte offsets <start, end>
 168  * @mapping:    address space structure to write
 169  * @start:      offset in bytes where the range starts
 170  * @end:        offset in bytes where the range ends
 171  * @sync_mode:  enable synchronous operation
 172  *
 173  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
 174  * opposed to a regular memory * cleansing writeback.  The difference between
 175  * these two operations is that if a dirty page/buffer is encountered, it must
 176  * be waited upon, and not just skipped over.
 177  */
 178 static int __filemap_fdatawrite_range(struct address_space *mapping,
 179         loff_t start, loff_t end, int sync_mode)
 180 {
 181         int ret;
 182         struct writeback_control wbc = {
 183                 .sync_mode = sync_mode,
 184                 .nr_to_write = mapping->nrpages * 2,
 185                 .start = start,
 186                 .end = end,
 187         };
 188
 189         if (!mapping_cap_writeback_dirty(mapping))
 190                 return 0;
 191
 192         ret = do_writepages(mapping, &wbc);
 193         return ret;
 194 }
 195
 196 static inline int __filemap_fdatawrite(struct address_space *mapping,
 197         int sync_mode)
 198 {
 199         return __filemap_fdatawrite_range(mapping, 0, 0, sync_mode);
 200 }
 201
 202 int filemap_fdatawrite(struct address_space *mapping)
 203 {
 204         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
 205 }
 206 EXPORT_SYMBOL(filemap_fdatawrite);
 207
 208 static int filemap_fdatawrite_range(struct address_space *mapping,
 209         loff_t start, loff_t end)
 210 {
 211         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
 212 }
 213
 214 /*
 215  * This is a mostly non-blocking flush.  Not suitable for data-integrity
 216  * purposes - I/O may not be started against all dirty pages.
 217  */
 218 int filemap_flush(struct address_space *mapping)
 219 {
 220         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
 221 }
 222 EXPORT_SYMBOL(filemap_flush);
 223
 224 /*
 225  * Wait for writeback to complete against pages indexed by start->end
 226  * inclusive
 227  */
 228 static int wait_on_page_writeback_range(struct address_space *mapping,
 229                                 pgoff_t start, pgoff_t end)
 230 {
 231         struct pagevec pvec;
 232         int nr_pages;
 233         int ret = 0;
 234         pgoff_t index;
 235
 236         if (end < start)
 237                 return 0;
 238
 239         pagevec_init(&pvec, 0);
 240         index = start;
 241         while ((index <= end) &&
 242                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
 243                         PAGECACHE_TAG_WRITEBACK,
 244                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
 245                 unsigned i;
 246
 247                 for (i = 0; i < nr_pages; i++) {
 248                         struct page *page = pvec.pages[i];
 249
 250                         /* until radix tree lookup accepts end_index */
 251                         if (page->index > end)
 252                                 continue;
 253
 254                         wait_on_page_writeback(page);
 255                         if (PageError(page))
 256                                 ret = -EIO;
 257                 }
 258                 pagevec_release(&pvec);
 259                 cond_resched();
 260         }
 261
 262         /* Check for outstanding write errors */
 263         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
 264                 ret = -ENOSPC;
 265         if (test_and_clear_bit(AS_EIO, &mapping->flags))
 266                 ret = -EIO;
 267
 268         return ret;
 269 }
 270
 271 /*
 272  * Write and wait upon all the pages in the passed range.  This is a "data
 273  * integrity" operation.  It waits upon in-flight writeout before starting and
 274  * waiting upon new writeout.  If there was an IO error, return it.
 275  *
 276  * We need to re-take i_sem during the generic_osync_inode list walk because
 277  * it is otherwise livelockable.
 278  */
 279 int sync_page_range(struct inode *inode, struct address_space *mapping,
 280                         loff_t pos, size_t count)
 281 {
 282         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
 283         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
 284         int ret;
 285
 286         if (!mapping_cap_writeback_dirty(mapping) || !count)
 287                 return 0;
 288         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
 289         if (ret == 0) {
 290                 down(&inode->i_sem);
 291                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
 292                 up(&inode->i_sem);
 293         }
 294         if (ret == 0)
 295                 ret = wait_on_page_writeback_range(mapping, start, end);
 296         return ret;
 297 }
 298 EXPORT_SYMBOL(sync_page_range);
 299
 300 /*
 301  * Note: Holding i_sem across sync_page_range_nolock is not a good idea
 302  * as it forces O_SYNC writers to different parts of the same file
 303  * to be serialised right until io completion.
 304  */
 305 int sync_page_range_nolock(struct inode *inode, struct address_space *mapping,
 306                         loff_t pos, size_t count)
 307 {
 308         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
 309         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
 310         int ret;
 311
 312         if (!mapping_cap_writeback_dirty(mapping) || !count)
 313                 return 0;
 314         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
 315         if (ret == 0)
 316                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
 317         if (ret == 0)
 318                 ret = wait_on_page_writeback_range(mapping, start, end);
 319         return ret;
 320 }
 321 EXPORT_SYMBOL(sync_page_range_nolock);
 322
 323 /**
 324  * filemap_fdatawait - walk the list of under-writeback pages of the given
 325  *     address space and wait for all of them.
 326  *
 327  * @mapping: address space structure to wait for
 328  */
 329 int filemap_fdatawait(struct address_space *mapping)
 330 {
 331         loff_t i_size = i_size_read(mapping->host);
 332
 333         if (i_size == 0)
 334                 return 0;
 335
 336         return wait_on_page_writeback_range(mapping, 0,
 337                                 (i_size - 1) >> PAGE_CACHE_SHIFT);
 338 }
 339 EXPORT_SYMBOL(filemap_fdatawait);
 340
 341 int filemap_write_and_wait(struct address_space *mapping)
 342 {
 343         int retval = 0;
 344
 345         if (mapping->nrpages) {
 346                 retval = filemap_fdatawrite(mapping);
 347                 if (retval == 0)
 348                         retval = filemap_fdatawait(mapping);
 349         }
 350         return retval;
 351 }
 352
 353 int filemap_write_and_wait_range(struct address_space *mapping,
 354                                  loff_t lstart, loff_t lend)
 355 {
 356         int retval = 0;
 357
 358         if (mapping->nrpages) {
 359                 retval = __filemap_fdatawrite_range(mapping, lstart, lend,
 360                                                     WB_SYNC_ALL);
 361                 if (retval == 0)
 362                         retval = wait_on_page_writeback_range(mapping,
 363                                                     lstart >> PAGE_CACHE_SHIFT,
 364                                                     lend >> PAGE_CACHE_SHIFT);
 365         }
 366         return retval;
 367 }
 368
 369 /*
 370  * This function is used to add newly allocated pagecache pages:
 371  * the page is new, so we can just run SetPageLocked() against it.
 372  * The other page state flags were set by rmqueue().
 373  *
 374  * This function does not add the page to the LRU.  The caller must do that.
 375  */
 376 int add_to_page_cache(struct page *page, struct address_space *mapping,
 377                 pgoff_t offset, int gfp_mask)
 378 {
 379         int error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
 380
 381         if (error == 0) {
 382                 write_lock_irq(&mapping->tree_lock);
 383                 error = radix_tree_insert(&mapping->page_tree, offset, page);
 384                 if (!error) {
 385                         page_cache_get(page);
 386                         SetPageLocked(page);
 387                         page->mapping = mapping;
 388                         page->index = offset;
 389                         mapping->nrpages++;
 390                         pagecache_acct(1);
 391                 }
 392                 write_unlock_irq(&mapping->tree_lock);
 393                 radix_tree_preload_end();
 394         }
 395         return error;
 396 }
 397
 398 EXPORT_SYMBOL(add_to_page_cache);
 399
 400 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
 401                                 pgoff_t offset, int gfp_mask)
 402 {
 403         int ret = add_to_page_cache(page, mapping, offset, gfp_mask);
 404         if (ret == 0)
 405                 lru_cache_add(page);
 406         return ret;
 407 }
 408
 409 /*
 410  * In order to wait for pages to become available there must be
 411  * waitqueues associated with pages. By using a hash table of
 412  * waitqueues where the bucket discipline is to maintain all
 413  * waiters on the same queue and wake all when any of the pages
 414  * become available, and for the woken contexts to check to be
 415  * sure the appropriate page became available, this saves space
 416  * at a cost of "thundering herd" phenomena during rare hash
 417  * collisions.
 418  */
 419 static wait_queue_head_t *page_waitqueue(struct page *page)
 420 {
 421         const struct zone *zone = page_zone(page);
 422
 423         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
 424 }
 425
 426 static inline void wake_up_page(struct page *page, int bit)
 427 {
 428         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
 429 }
 430
 431 void fastcall wait_on_page_bit(struct page *page, int bit_nr)
 432 {
 433         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
 434
 435         if (test_bit(bit_nr, &page->flags))
 436                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
 437                                                         TASK_UNINTERRUPTIBLE);
 438 }
 439 EXPORT_SYMBOL(wait_on_page_bit);
 440
 441 /**
 442  * unlock_page() - unlock a locked page
 443  *
 444  * @page: the page
 445  *
 446  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
 447  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
 448  * mechananism between PageLocked pages and PageWriteback pages is shared.
 449  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
 450  *
 451  * The first mb is necessary to safely close the critical section opened by the
 452  * TestSetPageLocked(), the second mb is necessary to enforce ordering between
 453  * the clear_bit and the read of the waitqueue (to avoid SMP races with a
 454  * parallel wait_on_page_locked()).
 455  */
 456 void fastcall unlock_page(struct page *page)
 457 {
 458         smp_mb__before_clear_bit();
 459         if (!TestClearPageLocked(page))
 460                 BUG();
 461         smp_mb__after_clear_bit();
 462         wake_up_page(page, PG_locked);
 463 }
 464 EXPORT_SYMBOL(unlock_page);
 465
 466 /*
 467  * End writeback against a page.
 468  */
 469 void end_page_writeback(struct page *page)
 470 {
 471         if (!TestClearPageReclaim(page) || rotate_reclaimable_page(page)) {
 472                 if (!test_clear_page_writeback(page))
 473                         BUG();
 474         }
 475         smp_mb__after_clear_bit();
 476         wake_up_page(page, PG_writeback);
 477 }
 478 EXPORT_SYMBOL(end_page_writeback);
 479
 480 /*
 481  * Get a lock on the page, assuming we need to sleep to get it.
 482  *
 483  * Ugly: running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
 484  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
 485  * chances are that on the second loop, the block layer's plug list is empty,
 486  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
 487  */
 488 void fastcall __lock_page(struct page *page)
 489 {
 490         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
 491
 492         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
 493                                                         TASK_UNINTERRUPTIBLE);
 494 }
 495 EXPORT_SYMBOL(__lock_page);
 496
 497 /*
 498  * a rather lightweight function, finding and getting a reference to a
 499  * hashed page atomically.
 500  */
 501 struct page * find_get_page(struct address_space *mapping, unsigned long offset)
 502 {
 503         struct page *page;
 504
 505         read_lock_irq(&mapping->tree_lock);
 506         page = radix_tree_lookup(&mapping->page_tree, offset);
 507         if (page)
 508                 page_cache_get(page);
 509         read_unlock_irq(&mapping->tree_lock);
 510         return page;
 511 }
 512
 513 EXPORT_SYMBOL(find_get_page);
 514
 515 /*
 516  * Same as above, but trylock it instead of incrementing the count.
 517  */
 518 struct page *find_trylock_page(struct address_space *mapping, unsigned long offset)
 519 {
 520         struct page *page;
 521
 522         read_lock_irq(&mapping->tree_lock);
 523         page = radix_tree_lookup(&mapping->page_tree, offset);
 524         if (page && TestSetPageLocked(page))
 525                 page = NULL;
 526         read_unlock_irq(&mapping->tree_lock);
 527         return page;
 528 }
 529
 530 EXPORT_SYMBOL(find_trylock_page);
 531
 532 /**
 533  * find_lock_page - locate, pin and lock a pagecache page
 534  *
 535  * @mapping: the address_space to search
 536  * @offset: the page index
 537  *
 538  * Locates the desired pagecache page, locks it, increments its reference
 539  * count and returns its address.
 540  *
 541  * Returns zero if the page was not present. find_lock_page() may sleep.
 542  */
 543 struct page *find_lock_page(struct address_space *mapping,
 544                                 unsigned long offset)
 545 {
 546         struct page *page;
 547
 548         read_lock_irq(&mapping->tree_lock);
 549 repeat:
 550         page = radix_tree_lookup(&mapping->page_tree, offset);
 551         if (page) {
 552                 page_cache_get(page);
 553                 if (TestSetPageLocked(page)) {
 554                         read_unlock_irq(&mapping->tree_lock);
 555                         lock_page(page);
 556                         read_lock_irq(&mapping->tree_lock);
 557
 558                         /* Has the page been truncated while we slept? */
 559                         if (page->mapping != mapping || page->index != offset) {
 560                                 unlock_page(page);
 561                                 page_cache_release(page);
 562                                 goto repeat;
 563                         }
 564                 }
 565         }
 566         read_unlock_irq(&mapping->tree_lock);
 567         return page;
 568 }
 569
 570 EXPORT_SYMBOL(find_lock_page);
 571
 572 /**
 573  * find_or_create_page - locate or add a pagecache page
 574  *
 575  * @mapping: the page's address_space
 576  * @index: the page's index into the mapping
 577  * @gfp_mask: page allocation mode
 578  *
 579  * Locates a page in the pagecache.  If the page is not present, a new page
 580  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
 581  * LRU list.  The returned page is locked and has its reference count
 582  * incremented.
 583  *
 584  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
 585  * allocation!
 586  *
 587  * find_or_create_page() returns the desired page's address, or zero on
 588  * memory exhaustion.
 589  */
 590 struct page *find_or_create_page(struct address_space *mapping,
 591                 unsigned long index, unsigned int gfp_mask)
 592 {
 593         struct page *page, *cached_page = NULL;
 594         int err;
 595 repeat:
 596         page = find_lock_page(mapping, index);
 597         if (!page) {
 598                 if (!cached_page) {
 599                         cached_page = alloc_page(gfp_mask);
 600                         if (!cached_page)
 601                                 return NULL;
 602                 }
 603                 err = add_to_page_cache_lru(cached_page, mapping,
 604                                         index, gfp_mask);
 605                 if (!err) {
 606                         page = cached_page;
 607                         cached_page = NULL;
 608                 } else if (err == -EEXIST)
 609                         goto repeat;
 610         }
 611         if (cached_page)
 612                 page_cache_release(cached_page);
 613         return page;
 614 }
 615
 616 EXPORT_SYMBOL(find_or_create_page);
 617
 618 /**
 619  * find_get_pages - gang pagecache lookup
 620  * @mapping:    The address_space to search
 621  * @start:      The starting page index
 622  * @nr_pages:   The maximum number of pages
 623  * @pages:      Where the resulting pages are placed
 624  *
 625  * find_get_pages() will search for and return a group of up to
 626  * @nr_pages pages in the mapping.  The pages are placed at @pages.
 627  * find_get_pages() takes a reference against the returned pages.
 628  *
 629  * The search returns a group of mapping-contiguous pages with ascending
 630  * indexes.  There may be holes in the indices due to not-present pages.
 631  *
 632  * find_get_pages() returns the number of pages which were found.
 633  */
 634 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
 635                             unsigned int nr_pages, struct page **pages)
 636 {
 637         unsigned int i;
 638         unsigned int ret;
 639
 640         read_lock_irq(&mapping->tree_lock);
 641         ret = radix_tree_gang_lookup(&mapping->page_tree,
 642                                 (void **)pages, start, nr_pages);
 643         for (i = 0; i < ret; i++)
 644                 page_cache_get(pages[i]);
 645         read_unlock_irq(&mapping->tree_lock);
 646         return ret;
 647 }
 648
 649 /*
 650  * Like find_get_pages, except we only return pages which are tagged with
 651  * `tag'.   We update *index to index the next page for the traversal.
 652  */
 653 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
 654                         int tag, unsigned int nr_pages, struct page **pages)
 655 {
 656         unsigned int i;
 657         unsigned int ret;
 658
 659         read_lock_irq(&mapping->tree_lock);
 660         ret = radix_tree_gang_lookup_tag(&mapping->page_tree,
 661                                 (void **)pages, *index, nr_pages, tag);
 662         for (i = 0; i < ret; i++)
 663                 page_cache_get(pages[i]);
 664         if (ret)
 665                 *index = pages[ret - 1]->index + 1;
 666         read_unlock_irq(&mapping->tree_lock);
 667         return ret;
 668 }
 669
 670 /*
 671  * Same as grab_cache_page, but do not wait if the page is unavailable.
 672  * This is intended for speculative data generators, where the data can
 673  * be regenerated if the page couldn't be grabbed.  This routine should
 674  * be safe to call while holding the lock for another page.
 675  *
 676  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
 677  * and deadlock against the caller's locked page.
 678  */
 679 struct page *
 680 grab_cache_page_nowait(struct address_space *mapping, unsigned long index)
 681 {
 682         struct page *page = find_get_page(mapping, index);
 683         unsigned int gfp_mask;
 684
 685         if (page) {
 686                 if (!TestSetPageLocked(page))
 687                         return page;
 688                 page_cache_release(page);
 689                 return NULL;
 690         }
 691         gfp_mask = mapping_gfp_mask(mapping) & ~__GFP_FS;
 692         page = alloc_pages(gfp_mask, 0);
 693         if (page && add_to_page_cache_lru(page, mapping, index, gfp_mask)) {
 694                 page_cache_release(page);
 695                 page = NULL;
 696         }
 697         return page;
 698 }
 699
 700 EXPORT_SYMBOL(grab_cache_page_nowait);
 701
 702 /*
 703  * This is a generic file read routine, and uses the
 704  * mapping->a_ops->readpage() function for the actual low-level
 705  * stuff.
 706  *
 707  * This is really ugly. But the goto's actually try to clarify some
 708  * of the logic when it comes to error handling etc.
 709  *
 710  * Note the struct file* is only passed for the use of readpage.  It may be
 711  * NULL.
 712  */
 713 void do_generic_mapping_read(struct address_space *mapping,
 714                              struct file_ra_state *_ra,
 715                              struct file *filp,
 716                              loff_t *ppos,
 717                              read_descriptor_t *desc,
 718                              read_actor_t actor)
 719 {
 720         struct inode *inode = mapping->host;
 721         unsigned long index;
 722         unsigned long end_index;
 723         unsigned long offset;
 724         unsigned long last_index;
 725         unsigned long next_index;
 726         unsigned long prev_index;
 727         loff_t isize;
 728         struct page *cached_page;
 729         int error;
 730         struct file_ra_state ra = *_ra;
 731
 732         cached_page = NULL;
 733         index = *ppos >> PAGE_CACHE_SHIFT;
 734         next_index = index;
 735         prev_index = ra.prev_page;
 736         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
 737         offset = *ppos & ~PAGE_CACHE_MASK;
 738
 739         isize = i_size_read(inode);
 740         if (!isize)
 741                 goto out;
 742
 743         end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 744         for (;;) {
 745                 struct page *page;
 746                 unsigned long nr, ret;
 747
 748                 /* nr is the maximum number of bytes to copy from this page */
 749                 nr = PAGE_CACHE_SIZE;
 750                 if (index >= end_index) {
 751                         if (index > end_index)
 752                                 goto out;
 753                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 754                         if (nr <= offset) {
 755                                 goto out;
 756                         }
 757                 }
 758                 nr = nr - offset;
 759
 760                 cond_resched();
 761                 if (index == next_index)
 762                         next_index = page_cache_readahead(mapping, &ra, filp,
 763                                         index, last_index - index);
 764
 765 find_page:
 766                 page = find_get_page(mapping, index);
 767                 if (unlikely(page == NULL)) {
 768                         handle_ra_miss(mapping, &ra, index);
 769                         goto no_cached_page;
 770                 }
 771                 if (!PageUptodate(page))
 772                         goto page_not_up_to_date;
 773 page_ok:
 774
 775                 /* If users can be writing to this page using arbitrary
 776                  * virtual addresses, take care about potential aliasing
 777                  * before reading the page on the kernel side.
 778                  */
 779                 if (mapping_writably_mapped(mapping))
 780                         flush_dcache_page(page);
 781
 782                 /*
 783                  * When (part of) the same page is read multiple times
 784                  * in succession, only mark it as accessed the first time.
 785                  */
 786                 if (prev_index != index)
 787                         mark_page_accessed(page);
 788                 prev_index = index;
 789
 790                 /*
 791                  * Ok, we have the page, and it's up-to-date, so
 792                  * now we can copy it to user space...
 793                  *
 794                  * The actor routine returns how many bytes were actually used..
 795                  * NOTE! This may not be the same as how much of a user buffer
 796                  * we filled up (we may be padding etc), so we can only update
 797                  * "pos" here (the actor routine has to update the user buffer
 798                  * pointers and the remaining count).
 799                  */
 800                 ret = actor(desc, page, offset, nr);
 801                 offset += ret;
 802                 index += offset >> PAGE_CACHE_SHIFT;
 803                 offset &= ~PAGE_CACHE_MASK;
 804
 805                 page_cache_release(page);
 806                 if (ret == nr && desc->count)
 807                         continue;
 808                 goto out;
 809
 810 page_not_up_to_date:
 811                 /* Get exclusive access to the page ... */
 812                 lock_page(page);
 813
 814                 /* Did it get unhashed before we got the lock? */
 815                 if (!page->mapping) {
 816                         unlock_page(page);
 817                         page_cache_release(page);
 818                         continue;
 819                 }
 820
 821                 /* Did somebody else fill it already? */
 822                 if (PageUptodate(page)) {
 823                         unlock_page(page);
 824                         goto page_ok;
 825                 }
 826
 827 readpage:
 828                 /* Start the actual read. The read will unlock the page. */
 829                 error = mapping->a_ops->readpage(filp, page);
 830
 831                 if (unlikely(error))
 832                         goto readpage_error;
 833
 834                 if (!PageUptodate(page)) {
 835                         lock_page(page);
 836                         if (!PageUptodate(page)) {
 837                                 if (page->mapping == NULL) {
 838                                         /*
 839                                          * invalidate_inode_pages got it
 840                                          */
 841                                         unlock_page(page);
 842                                         page_cache_release(page);
 843                                         goto find_page;
 844                                 }
 845                                 unlock_page(page);
 846                                 error = -EIO;
 847                                 goto readpage_error;
 848                         }
 849                         unlock_page(page);
 850                 }
 851
 852                 /*
 853                  * i_size must be checked after we have done ->readpage.
 854                  *
 855                  * Checking i_size after the readpage allows us to calculate
 856                  * the correct value for "nr", which means the zero-filled
 857                  * part of the page is not copied back to userspace (unless
 858                  * another truncate extends the file - this is desired though).
 859                  */
 860                 isize = i_size_read(inode);
 861                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
 862                 if (unlikely(!isize || index > end_index)) {
 863                         page_cache_release(page);
 864                         goto out;
 865                 }
 866
 867                 /* nr is the maximum number of bytes to copy from this page */
 868                 nr = PAGE_CACHE_SIZE;
 869                 if (index == end_index) {
 870                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
 871                         if (nr <= offset) {
 872                                 page_cache_release(page);
 873                                 goto out;
 874                         }
 875                 }
 876                 nr = nr - offset;
 877                 goto page_ok;
 878
 879 readpage_error:
 880                 /* UHHUH! A synchronous read error occurred. Report it */
 881                 desc->error = error;
 882                 page_cache_release(page);
 883                 goto out;
 884
 885 no_cached_page:
 886                 /*
 887                  * Ok, it wasn't cached, so we need to create a new
 888                  * page..
 889                  */
 890                 if (!cached_page) {
 891                         cached_page = page_cache_alloc_cold(mapping);
 892                         if (!cached_page) {
 893                                 desc->error = -ENOMEM;
 894                                 goto out;
 895                         }
 896                 }
 897                 error = add_to_page_cache_lru(cached_page, mapping,
 898                                                 index, GFP_KERNEL);
 899                 if (error) {
 900                         if (error == -EEXIST)
 901                                 goto find_page;
 902                         desc->error = error;
 903                         goto out;
 904                 }
 905                 page = cached_page;
 906                 cached_page = NULL;
 907                 goto readpage;
 908         }
 909
 910 out:
 911         *_ra = ra;
 912
 913         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
 914         if (cached_page)
 915                 page_cache_release(cached_page);
 916         if (filp)
 917                 file_accessed(filp);
 918 }
 919
 920 EXPORT_SYMBOL(do_generic_mapping_read);
 921
 922 int file_read_actor(read_descriptor_t *desc, struct page *page,
 923                         unsigned long offset, unsigned long size)
 924 {
 925         char *kaddr;
 926         unsigned long left, count = desc->count;
 927
 928         if (size > count)
 929                 size = count;
 930
 931         /*
 932          * Faults on the destination of a read are common, so do it before
 933          * taking the kmap.
 934          */
 935         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
 936                 kaddr = kmap_atomic(page, KM_USER0);
 937                 left = __copy_to_user_inatomic(desc->arg.buf,
 938                                                 kaddr + offset, size);
 939                 kunmap_atomic(kaddr, KM_USER0);
 940                 if (left == 0)
 941                         goto success;
 942         }
 943
 944         /* Do it the slow way */
 945         kaddr = kmap(page);
 946         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
 947         kunmap(page);
 948
 949         if (left) {
 950                 size -= left;
 951                 desc->error = -EFAULT;
 952         }
 953 success:
 954         desc->count = count - size;
 955         desc->written += size;
 956         desc->arg.buf += size;
 957         return size;
 958 }
 959
 960 /*
 961  * This is the "read()" routine for all filesystems
 962  * that can use the page cache directly.
 963  */
 964 ssize_t
 965 __generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
 966                 unsigned long nr_segs, loff_t *ppos)
 967 {
 968         struct file *filp = iocb->ki_filp;
 969         ssize_t retval;
 970         unsigned long seg;
 971         size_t count;
 972
 973         count = 0;
 974         for (seg = 0; seg < nr_segs; seg++) {
 975                 const struct iovec *iv = &iov[seg];
 976
 977                 /*
 978                  * If any segment has a negative length, or the cumulative
 979                  * length ever wraps negative then return -EINVAL.
 980                  */
 981                 count += iv->iov_len;
 982                 if (unlikely((ssize_t)(count|iv->iov_len) < 0))
 983                         return -EINVAL;
 984                 if (access_ok(VERIFY_WRITE, iv->iov_base, iv->iov_len))
 985                         continue;
 986                 if (seg == 0)
 987                         return -EFAULT;
 988                 nr_segs = seg;
 989                 count -= iv->iov_len;   /* This segment is no good */
 990                 break;
 991         }
 992
 993         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
 994         if (filp->f_flags & O_DIRECT) {
 995                 loff_t pos = *ppos, size;
 996                 struct address_space *mapping;
 997                 struct inode *inode;
 998
 999                 mapping = filp->f_mapping;
1000                 inode = mapping->host;
1001                 retval = 0;
1002                 if (!count)
1003                         goto out; /* skip atime */
1004                 size = i_size_read(inode);
1005                 if (pos < size) {
1006                         retval = generic_file_direct_IO(READ, iocb,
1007                                                 iov, pos, nr_segs);
1008                         if (retval > 0 && !is_sync_kiocb(iocb))
1009                                 retval = -EIOCBQUEUED;
1010                         if (retval > 0)
1011                                 *ppos = pos + retval;
1012                 }
1013                 file_accessed(filp);
1014                 goto out;
1015         }
1016
1017         retval = 0;
1018         if (count) {
1019                 for (seg = 0; seg < nr_segs; seg++) {
1020                         read_descriptor_t desc;
1021
1022                         desc.written = 0;
1023                         desc.arg.buf = iov[seg].iov_base;
1024                         desc.count = iov[seg].iov_len;
1025                         if (desc.count == 0)
1026                                 continue;
1027                         desc.error = 0;
1028                         do_generic_file_read(filp,ppos,&desc,file_read_actor);
1029                         retval += desc.written;
1030                         if (!retval) {
1031                                 retval = desc.error;
1032                                 break;
1033                         }
1034                 }
1035         }
1036 out:
1037         return retval;
1038 }
1039
1040 EXPORT_SYMBOL(__generic_file_aio_read);
1041
1042 ssize_t
1043 generic_file_aio_read(struct kiocb *iocb, char __user *buf, size_t count, loff_t pos)
1044 {
1045         struct iovec local_iov = { .iov_base = buf, .iov_len = count };
1046
1047         BUG_ON(iocb->ki_pos != pos);
1048         return __generic_file_aio_read(iocb, &local_iov, 1, &iocb->ki_pos);
1049 }
1050
1051 EXPORT_SYMBOL(generic_file_aio_read);
1052
1053 ssize_t
1054 generic_file_read(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
1055 {
1056         struct iovec local_iov = { .iov_base = buf, .iov_len = count };
1057         struct kiocb kiocb;
1058         ssize_t ret;
1059
1060         init_sync_kiocb(&kiocb, filp);
1061         ret = __generic_file_aio_read(&kiocb, &local_iov, 1, ppos);
1062         if (-EIOCBQUEUED == ret)
1063                 ret = wait_on_sync_kiocb(&kiocb);
1064         return ret;
1065 }
1066
1067 EXPORT_SYMBOL(generic_file_read);
1068
1069 int file_send_actor(read_descriptor_t * desc, struct page *page, unsigned long offset, unsigned long size)
1070 {
1071         ssize_t written;
1072         unsigned long count = desc->count;
1073         struct file *file = desc->arg.data;
1074
1075         if (size > count)
1076                 size = count;
1077
1078         written = file->f_op->sendpage(file, page, offset,
1079                                        size, &file->f_pos, size<count);
1080         if (written < 0) {
1081                 desc->error = written;
1082                 written = 0;
1083         }
1084         desc->count = count - written;
1085         desc->written += written;
1086         return written;
1087 }
1088
1089 ssize_t generic_file_sendfile(struct file *in_file, loff_t *ppos,
1090                          size_t count, read_actor_t actor, void *target)
1091 {
1092         read_descriptor_t desc;
1093
1094         if (!count)
1095                 return 0;
1096
1097         desc.written = 0;
1098         desc.count = count;
1099         desc.arg.data = target;
1100         desc.error = 0;
1101
1102         do_generic_file_read(in_file, ppos, &desc, actor);
1103         if (desc.written)
1104                 return desc.written;
1105         return desc.error;
1106 }
1107
1108 EXPORT_SYMBOL(generic_file_sendfile);
1109
1110 static ssize_t
1111 do_readahead(struct address_space *mapping, struct file *filp,
1112              unsigned long index, unsigned long nr)
1113 {
1114         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1115                 return -EINVAL;
1116
1117         force_page_cache_readahead(mapping, filp, index,
1118                                         max_sane_readahead(nr));
1119         return 0;
1120 }
1121
1122 asmlinkage ssize_t sys_readahead(int fd, loff_t offset, size_t count)
1123 {
1124         ssize_t ret;
1125         struct file *file;
1126
1127         ret = -EBADF;
1128         file = fget(fd);
1129         if (file) {
1130                 if (file->f_mode & FMODE_READ) {
1131                         struct address_space *mapping = file->f_mapping;
1132                         unsigned long start = offset >> PAGE_CACHE_SHIFT;
1133                         unsigned long end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1134                         unsigned long len = end - start + 1;
1135                         ret = do_readahead(mapping, file, start, len);
1136                 }
1137                 fput(file);
1138         }
1139         return ret;
1140 }
1141
1142 #ifdef CONFIG_MMU
1143 /*
1144  * This adds the requested page to the page cache if it isn't already there,
1145  * and schedules an I/O to read in its contents from disk.
1146  */
1147 static int FASTCALL(page_cache_read(struct file * file, unsigned long offset));
1148 static int fastcall page_cache_read(struct file * file, unsigned long offset)
1149 {
1150         struct address_space *mapping = file->f_mapping;
1151         struct page *page;
1152         int error;
1153
1154         page = page_cache_alloc_cold(mapping);
1155         if (!page)
1156                 return -ENOMEM;
1157
1158         error = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1159         if (!error) {
1160                 error = mapping->a_ops->readpage(file, page);
1161                 page_cache_release(page);
1162                 return error;
1163         }
1164
1165         /*
1166          * We arrive here in the unlikely event that someone
1167          * raced with us and added our page to the cache first
1168          * or we are out of memory for radix-tree nodes.
1169          */
1170         page_cache_release(page);
1171         return error == -EEXIST ? 0 : error;
1172 }
1173
1174 #define MMAP_LOTSAMISS  (100)
1175
1176 /*
1177  * filemap_nopage() is invoked via the vma operations vector for a
1178  * mapped memory region to read in file data during a page fault.
1179  *
1180  * The goto's are kind of ugly, but this streamlines the normal case of having
1181  * it in the page cache, and handles the special cases reasonably without
1182  * having a lot of duplicated code.
1183  */
1184 struct page *filemap_nopage(struct vm_area_struct *area,
1185                                 unsigned long address, int *type)
1186 {
1187         int error;
1188         struct file *file = area->vm_file;
1189         struct address_space *mapping = file->f_mapping;
1190         struct file_ra_state *ra = &file->f_ra;
1191         struct inode *inode = mapping->host;
1192         struct page *page;
1193         unsigned long size, pgoff;
1194         int did_readaround = 0, majmin = VM_FAULT_MINOR;
1195
1196         pgoff = ((address-area->vm_start) >> PAGE_CACHE_SHIFT) + area->vm_pgoff;
1197
1198 retry_all:
1199         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1200         if (pgoff >= size)
1201                 goto outside_data_content;
1202
1203         /* If we don't want any read-ahead, don't bother */
1204         if (VM_RandomReadHint(area))
1205                 goto no_cached_page;
1206
1207         /*
1208          * The readahead code wants to be told about each and every page
1209          * so it can build and shrink its windows appropriately
1210          *
1211          * For sequential accesses, we use the generic readahead logic.
1212          */
1213         if (VM_SequentialReadHint(area))
1214                 page_cache_readahead(mapping, ra, file, pgoff, 1);
1215
1216         /*
1217          * Do we have something in the page cache already?
1218          */
1219 retry_find:
1220         page = find_get_page(mapping, pgoff);
1221         if (!page) {
1222                 unsigned long ra_pages;
1223
1224                 if (VM_SequentialReadHint(area)) {
1225                         handle_ra_miss(mapping, ra, pgoff);
1226                         goto no_cached_page;
1227                 }
1228                 ra->mmap_miss++;
1229
1230                 /*
1231                  * Do we miss much more than hit in this file? If so,
1232                  * stop bothering with read-ahead. It will only hurt.
1233                  */
1234                 if (ra->mmap_miss > ra->mmap_hit + MMAP_LOTSAMISS)
1235                         goto no_cached_page;
1236
1237                 /*
1238                  * To keep the pgmajfault counter straight, we need to
1239                  * check did_readaround, as this is an inner loop.
1240                  */
1241                 if (!did_readaround) {
1242                         majmin = VM_FAULT_MAJOR;
1243                         inc_page_state(pgmajfault);
1244                 }
1245                 did_readaround = 1;
1246                 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1247                 if (ra_pages) {
1248                         pgoff_t start = 0;
1249
1250                         if (pgoff > ra_pages / 2)
1251                                 start = pgoff - ra_pages / 2;
1252                         do_page_cache_readahead(mapping, file, start, ra_pages);
1253                 }
1254                 page = find_get_page(mapping, pgoff);
1255                 if (!page)
1256                         goto no_cached_page;
1257         }
1258
1259         if (!did_readaround)
1260                 ra->mmap_hit++;
1261
1262         /*
1263          * Ok, found a page in the page cache, now we need to check
1264          * that it's up-to-date.
1265          */
1266         if (!PageUptodate(page))
1267                 goto page_not_uptodate;
1268
1269 success:
1270         /*
1271          * Found the page and have a reference on it.
1272          */
1273         mark_page_accessed(page);
1274         if (type)
1275                 *type = majmin;
1276         return page;
1277
1278 outside_data_content:
1279         /*
1280          * An external ptracer can access pages that normally aren't
1281          * accessible..
1282          */
1283         if (area->vm_mm == current->mm)
1284                 return NULL;
1285         /* Fall through to the non-read-ahead case */
1286 no_cached_page:
1287         /*
1288          * We're only likely to ever get here if MADV_RANDOM is in
1289          * effect.
1290          */
1291         error = page_cache_read(file, pgoff);
1292         grab_swap_token();
1293
1294         /*
1295          * The page we want has now been added to the page cache.
1296          * In the unlikely event that someone removed it in the
1297          * meantime, we'll just come back here and read it again.
1298          */
1299         if (error >= 0)
1300                 goto retry_find;
1301
1302         /*
1303          * An error return from page_cache_read can result if the
1304          * system is low on memory, or a problem occurs while trying
1305          * to schedule I/O.
1306          */
1307         if (error == -ENOMEM)
1308                 return NOPAGE_OOM;
1309         return NULL;
1310
1311 page_not_uptodate:
1312         if (!did_readaround) {
1313                 majmin = VM_FAULT_MAJOR;
1314                 inc_page_state(pgmajfault);
1315         }
1316         lock_page(page);
1317
1318         /* Did it get unhashed while we waited for it? */
1319         if (!page->mapping) {
1320                 unlock_page(page);
1321                 page_cache_release(page);
1322                 goto retry_all;
1323         }
1324
1325         /* Did somebody else get it up-to-date? */
1326         if (PageUptodate(page)) {
1327                 unlock_page(page);
1328                 goto success;
1329         }
1330
1331         if (!mapping->a_ops->readpage(file, page)) {
1332                 wait_on_page_locked(page);
1333                 if (PageUptodate(page))
1334                         goto success;
1335         }
1336
1337         /*
1338          * Umm, take care of errors if the page isn't up-to-date.
1339          * Try to re-read it _once_. We do this synchronously,
1340          * because there really aren't any performance issues here
1341          * and we need to check for errors.
1342          */
1343         lock_page(page);
1344
1345         /* Somebody truncated the page on us? */
1346         if (!page->mapping) {
1347                 unlock_page(page);
1348                 page_cache_release(page);
1349                 goto retry_all;
1350         }
1351
1352         /* Somebody else successfully read it in? */
1353         if (PageUptodate(page)) {
1354                 unlock_page(page);
1355                 goto success;
1356         }
1357         ClearPageError(page);
1358         if (!mapping->a_ops->readpage(file, page)) {
1359                 wait_on_page_locked(page);
1360                 if (PageUptodate(page))
1361                         goto success;
1362         }
1363
1364         /*
1365          * Things didn't work out. Return zero to tell the
1366          * mm layer so, possibly freeing the page cache page first.
1367          */
1368         page_cache_release(page);
1369         return NULL;
1370 }
1371
1372 EXPORT_SYMBOL(filemap_nopage);
1373
1374 static struct page * filemap_getpage(struct file *file, unsigned long pgoff,
1375                                         int nonblock)
1376 {
1377         struct address_space *mapping = file->f_mapping;
1378         struct page *page;
1379         int error;
1380
1381         /*
1382          * Do we have something in the page cache already?
1383          */
1384 retry_find:
1385         page = find_get_page(mapping, pgoff);
1386         if (!page) {
1387                 if (nonblock)
1388                         return NULL;
1389                 goto no_cached_page;
1390         }
1391
1392         /*
1393          * Ok, found a page in the page cache, now we need to check
1394          * that it's up-to-date.
1395          */
1396         if (!PageUptodate(page)) {
1397                 if (nonblock) {
1398                         page_cache_release(page);
1399                         return NULL;
1400                 }
1401                 goto page_not_uptodate;
1402         }
1403
1404 success:
1405         /*
1406          * Found the page and have a reference on it.
1407          */
1408         mark_page_accessed(page);
1409         return page;
1410
1411 no_cached_page:
1412         error = page_cache_read(file, pgoff);
1413
1414         /*
1415          * The page we want has now been added to the page cache.
1416          * In the unlikely event that someone removed it in the
1417          * meantime, we'll just come back here and read it again.
1418          */
1419         if (error >= 0)
1420                 goto retry_find;
1421
1422         /*
1423          * An error return from page_cache_read can result if the
1424          * system is low on memory, or a problem occurs while trying
1425          * to schedule I/O.
1426          */
1427         return NULL;
1428
1429 page_not_uptodate:
1430         lock_page(page);
1431
1432         /* Did it get unhashed while we waited for it? */
1433         if (!page->mapping) {
1434                 unlock_page(page);
1435                 goto err;
1436         }
1437
1438         /* Did somebody else get it up-to-date? */
1439         if (PageUptodate(page)) {
1440                 unlock_page(page);
1441                 goto success;
1442         }
1443
1444         if (!mapping->a_ops->readpage(file, page)) {
1445                 wait_on_page_locked(page);
1446                 if (PageUptodate(page))
1447                         goto success;
1448         }
1449
1450         /*
1451          * Umm, take care of errors if the page isn't up-to-date.
1452          * Try to re-read it _once_. We do this synchronously,
1453          * because there really aren't any performance issues here
1454          * and we need to check for errors.
1455          */
1456         lock_page(page);
1457
1458         /* Somebody truncated the page on us? */
1459         if (!page->mapping) {
1460                 unlock_page(page);
1461                 goto err;
1462         }
1463         /* Somebody else successfully read it in? */
1464         if (PageUptodate(page)) {
1465                 unlock_page(page);
1466                 goto success;
1467         }
1468
1469         ClearPageError(page);
1470         if (!mapping->a_ops->readpage(file, page)) {
1471                 wait_on_page_locked(page);
1472                 if (PageUptodate(page))
1473                         goto success;
1474         }
1475
1476         /*
1477          * Things didn't work out. Return zero to tell the
1478          * mm layer so, possibly freeing the page cache page first.
1479          */
1480 err:
1481         page_cache_release(page);
1482
1483         return NULL;
1484 }
1485
1486 int filemap_populate(struct vm_area_struct *vma, unsigned long addr,
1487                 unsigned long len, pgprot_t prot, unsigned long pgoff,
1488                 int nonblock)
1489 {
1490         struct file *file = vma->vm_file;
1491         struct address_space *mapping = file->f_mapping;
1492         struct inode *inode = mapping->host;
1493         unsigned long size;
1494         struct mm_struct *mm = vma->vm_mm;
1495         struct page *page;
1496         int err;
1497
1498         if (!nonblock)
1499                 force_page_cache_readahead(mapping, vma->vm_file,
1500                                         pgoff, len >> PAGE_CACHE_SHIFT);
1501
1502 repeat:
1503         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1504         if (pgoff + (len >> PAGE_CACHE_SHIFT) > size)
1505                 return -EINVAL;
1506
1507         page = filemap_getpage(file, pgoff, nonblock);
1508         if (!page && !nonblock)
1509                 return -ENOMEM;
1510         if (page) {
1511                 err = install_page(mm, vma, addr, page, prot);
1512                 if (err) {
1513                         page_cache_release(page);
1514                         return err;
1515                 }
1516         } else {
1517                 err = install_file_pte(mm, vma, addr, pgoff, prot);
1518                 if (err)
1519                         return err;
1520         }
1521
1522         len -= PAGE_SIZE;
1523         addr += PAGE_SIZE;
1524         pgoff++;
1525         if (len)
1526                 goto repeat;
1527
1528         return 0;
1529 }
1530
1531 struct vm_operations_struct generic_file_vm_ops = {
1532         .nopage         = filemap_nopage,
1533         .populate       = filemap_populate,
1534 };
1535
1536 /* This is used for a general mmap of a disk file */
1537
1538 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1539 {
1540         struct address_space *mapping = file->f_mapping;
1541
1542         if (!mapping->a_ops->readpage)
1543                 return -ENOEXEC;
1544         file_accessed(file);
1545         vma->vm_ops = &generic_file_vm_ops;
1546         return 0;
1547 }
1548 EXPORT_SYMBOL(filemap_populate);
1549
1550 /*
1551  * This is for filesystems which do not implement ->writepage.
1552  */
1553 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1554 {
1555         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1556                 return -EINVAL;
1557         return generic_file_mmap(file, vma);
1558 }
1559 #else
1560 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1561 {
1562         return -ENOSYS;
1563 }
1564 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1565 {
1566         return -ENOSYS;
1567 }
1568 #endif /* CONFIG_MMU */
1569
1570 EXPORT_SYMBOL(generic_file_mmap);
1571 EXPORT_SYMBOL(generic_file_readonly_mmap);
1572
1573 static inline struct page *__read_cache_page(struct address_space *mapping,
1574                                 unsigned long index,
1575                                 int (*filler)(void *,struct page*),
1576                                 void *data)
1577 {
1578         struct page *page, *cached_page = NULL;
1579         int err;
1580 repeat:
1581         page = find_get_page(mapping, index);
1582         if (!page) {
1583                 if (!cached_page) {
1584                         cached_page = page_cache_alloc_cold(mapping);
1585                         if (!cached_page)
1586                                 return ERR_PTR(-ENOMEM);
1587                 }
1588                 err = add_to_page_cache_lru(cached_page, mapping,
1589                                         index, GFP_KERNEL);
1590                 if (err == -EEXIST)
1591                         goto repeat;
1592                 if (err < 0) {
1593                         /* Presumably ENOMEM for radix tree node */
1594                         page_cache_release(cached_page);
1595                         return ERR_PTR(err);
1596                 }
1597                 page = cached_page;
1598                 cached_page = NULL;
1599                 err = filler(data, page);
1600                 if (err < 0) {
1601                         page_cache_release(page);
1602                         page = ERR_PTR(err);
1603                 }
1604         }
1605         if (cached_page)
1606                 page_cache_release(cached_page);
1607         return page;
1608 }
1609
1610 /*
1611  * Read into the page cache. If a page already exists,
1612  * and PageUptodate() is not set, try to fill the page.
1613  */
1614 struct page *read_cache_page(struct address_space *mapping,
1615                                 unsigned long index,
1616                                 int (*filler)(void *,struct page*),
1617                                 void *data)
1618 {
1619         struct page *page;
1620         int err;
1621
1622 retry:
1623         page = __read_cache_page(mapping, index, filler, data);
1624         if (IS_ERR(page))
1625                 goto out;
1626         mark_page_accessed(page);
1627         if (PageUptodate(page))
1628                 goto out;
1629
1630         lock_page(page);
1631         if (!page->mapping) {
1632                 unlock_page(page);
1633                 page_cache_release(page);
1634                 goto retry;
1635         }
1636         if (PageUptodate(page)) {
1637                 unlock_page(page);
1638                 goto out;
1639         }
1640         err = filler(data, page);
1641         if (err < 0) {
1642                 page_cache_release(page);
1643                 page = ERR_PTR(err);
1644         }
1645  out:
1646         return page;
1647 }
1648
1649 EXPORT_SYMBOL(read_cache_page);
1650
1651 /*
1652  * If the page was newly created, increment its refcount and add it to the
1653  * caller's lru-buffering pagevec.  This function is specifically for
1654  * generic_file_write().
1655  */
1656 static inline struct page *
1657 __grab_cache_page(struct address_space *mapping, unsigned long index,
1658                         struct page **cached_page, struct pagevec *lru_pvec)
1659 {
1660         int err;
1661         struct page *page;
1662 repeat:
1663         page = find_lock_page(mapping, index);
1664         if (!page) {
1665                 if (!*cached_page) {
1666                         *cached_page = page_cache_alloc(mapping);
1667                         if (!*cached_page)
1668                                 return NULL;
1669                 }
1670                 err = add_to_page_cache(*cached_page, mapping,
1671                                         index, GFP_KERNEL);
1672                 if (err == -EEXIST)
1673                         goto repeat;
1674                 if (err == 0) {
1675                         page = *cached_page;
1676                         page_cache_get(page);
1677                         if (!pagevec_add(lru_pvec, page))
1678                                 __pagevec_lru_add(lru_pvec);
1679                         *cached_page = NULL;
1680                 }
1681         }
1682         return page;
1683 }
1684
1685 /*
1686  * The logic we want is
1687  *
1688  *      if suid or (sgid and xgrp)
1689  *              remove privs
1690  */
1691 int remove_suid(struct dentry *dentry)
1692 {
1693         mode_t mode = dentry->d_inode->i_mode;
1694         int kill = 0;
1695         int result = 0;
1696
1697         /* suid always must be killed */
1698         if (unlikely(mode & S_ISUID))
1699                 kill = ATTR_KILL_SUID;
1700
1701         /*
1702          * sgid without any exec bits is just a mandatory locking mark; leave
1703          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1704          */
1705         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1706                 kill |= ATTR_KILL_SGID;
1707
1708         if (unlikely(kill && !capable(CAP_FSETID))) {
1709                 struct iattr newattrs;
1710
1711                 newattrs.ia_valid = ATTR_FORCE | kill;
1712                 result = notify_change(dentry, &newattrs);
1713         }
1714         return result;
1715 }
1716 EXPORT_SYMBOL(remove_suid);
1717
1718 size_t
1719 __filemap_copy_from_user_iovec(char *vaddr,
1720                         const struct iovec *iov, size_t base, size_t bytes)
1721 {
1722         size_t copied = 0, left = 0;
1723
1724         while (bytes) {
1725                 char __user *buf = iov->iov_base + base;
1726                 int copy = min(bytes, iov->iov_len - base);
1727
1728                 base = 0;
1729                 left = __copy_from_user_inatomic(vaddr, buf, copy);
1730                 copied += copy;
1731                 bytes -= copy;
1732                 vaddr += copy;
1733                 iov++;
1734
1735                 if (unlikely(left)) {
1736                         /* zero the rest of the target like __copy_from_user */
1737                         if (bytes)
1738                                 memset(vaddr, 0, bytes);
1739                         break;
1740                 }
1741         }
1742         return copied - left;
1743 }
1744
1745 /*
1746  * Performs necessary checks before doing a write
1747  *
1748  * Can adjust writing position aor amount of bytes to write.
1749  * Returns appropriate error code that caller should return or
1750  * zero in case that write should be allowed.
1751  */
1752 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1753 {
1754         struct inode *inode = file->f_mapping->host;
1755         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1756
1757         if (unlikely(*pos < 0))
1758                 return -EINVAL;
1759
1760         if (!isblk) {
1761                 /* FIXME: this is for backwards compatibility with 2.4 */
1762                 if (file->f_flags & O_APPEND)
1763                         *pos = i_size_read(inode);
1764
1765                 if (limit != RLIM_INFINITY) {
1766                         if (*pos >= limit) {
1767                                 send_sig(SIGXFSZ, current, 0);
1768                                 return -EFBIG;
1769                         }
1770                         if (*count > limit - (typeof(limit))*pos) {
1771                                 *count = limit - (typeof(limit))*pos;
1772                         }
1773                 }
1774         }
1775
1776         /*
1777          * LFS rule
1778          */
1779         if (unlikely(*pos + *count > MAX_NON_LFS &&
1780                                 !(file->f_flags & O_LARGEFILE))) {
1781                 if (*pos >= MAX_NON_LFS) {
1782                         send_sig(SIGXFSZ, current, 0);
1783                         return -EFBIG;
1784                 }
1785                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
1786                         *count = MAX_NON_LFS - (unsigned long)*pos;
1787                 }
1788         }
1789
1790         /*
1791          * Are we about to exceed the fs block limit ?
1792          *
1793          * If we have written data it becomes a short write.  If we have
1794          * exceeded without writing data we send a signal and return EFBIG.
1795          * Linus frestrict idea will clean these up nicely..
1796          */
1797         if (likely(!isblk)) {
1798                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
1799                         if (*count || *pos > inode->i_sb->s_maxbytes) {
1800                                 send_sig(SIGXFSZ, current, 0);
1801                                 return -EFBIG;
1802                         }
1803                         /* zero-length writes at ->s_maxbytes are OK */
1804                 }
1805
1806                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
1807                         *count = inode->i_sb->s_maxbytes - *pos;
1808         } else {
1809                 loff_t isize;
1810                 if (bdev_read_only(I_BDEV(inode)))
1811                         return -EPERM;
1812                 isize = i_size_read(inode);
1813                 if (*pos >= isize) {
1814                         if (*count || *pos > isize)
1815                                 return -ENOSPC;
1816                 }
1817
1818                 if (*pos + *count > isize)
1819                         *count = isize - *pos;
1820         }
1821         return 0;
1822 }
1823 EXPORT_SYMBOL(generic_write_checks);
1824
1825 ssize_t
1826 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
1827                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
1828                 size_t count, size_t ocount)
1829 {
1830         struct file     *file = iocb->ki_filp;
1831         struct address_space *mapping = file->f_mapping;
1832         struct inode    *inode = mapping->host;
1833         ssize_t         written;
1834
1835         if (count != ocount)
1836                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
1837
1838         written = generic_file_direct_IO(WRITE, iocb, iov, pos, *nr_segs);
1839         if (written > 0) {
1840                 loff_t end = pos + written;
1841                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
1842                         i_size_write(inode,  end);
1843                         mark_inode_dirty(inode);
1844                 }
1845                 *ppos = end;
1846         }
1847
1848         /*
1849          * Sync the fs metadata but not the minor inode changes and
1850          * of course not the data as we did direct DMA for the IO.
1851          * i_sem is held, which protects generic_osync_inode() from
1852          * livelocking.
1853          */
1854         if (written >= 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
1855                 int err = generic_osync_inode(inode, mapping, OSYNC_METADATA);
1856                 if (err < 0)
1857                         written = err;
1858         }
1859         if (written == count && !is_sync_kiocb(iocb))
1860                 written = -EIOCBQUEUED;
1861         return written;
1862 }
1863 EXPORT_SYMBOL(generic_file_direct_write);
1864
1865 ssize_t
1866 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
1867                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
1868                 size_t count, ssize_t written)
1869 {
1870         struct file *file = iocb->ki_filp;
1871         struct address_space * mapping = file->f_mapping;
1872         struct address_space_operations *a_ops = mapping->a_ops;
1873         struct inode    *inode = mapping->host;
1874         long            status = 0;
1875         struct page     *page;
1876         struct page     *cached_page = NULL;
1877         size_t          bytes;
1878         struct pagevec  lru_pvec;
1879         const struct iovec *cur_iov = iov; /* current iovec */
1880         size_t          iov_base = 0;      /* offset in the current iovec */
1881         char __user     *buf;
1882
1883         pagevec_init(&lru_pvec, 0);
1884
1885         /*
1886          * handle partial DIO write.  Adjust cur_iov if needed.
1887          */
1888         if (likely(nr_segs == 1))
1889                 buf = iov->iov_base + written;
1890         else {
1891                 filemap_set_next_iovec(&cur_iov, &iov_base, written);
1892                 buf = cur_iov->iov_base + iov_base;
1893         }
1894
1895         do {
1896                 unsigned long index;
1897                 unsigned long offset;
1898                 unsigned long maxlen;
1899                 size_t copied;
1900
1901                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
1902                 index = pos >> PAGE_CACHE_SHIFT;
1903                 bytes = PAGE_CACHE_SIZE - offset;
1904                 if (bytes > count)
1905                         bytes = count;
1906
1907                 /*
1908                  * Bring in the user page that we will copy from _first_.
1909                  * Otherwise there's a nasty deadlock on copying from the
1910                  * same page as we're writing to, without it being marked
1911                  * up-to-date.
1912                  */
1913                 maxlen = cur_iov->iov_len - iov_base;
1914                 if (maxlen > bytes)
1915                         maxlen = bytes;
1916                 fault_in_pages_readable(buf, maxlen);
1917
1918                 page = __grab_cache_page(mapping,index,&cached_page,&lru_pvec);
1919                 if (!page) {
1920                         status = -ENOMEM;
1921                         break;
1922                 }
1923
1924                 status = a_ops->prepare_write(file, page, offset, offset+bytes);
1925                 if (unlikely(status)) {
1926                         loff_t isize = i_size_read(inode);
1927                         /*
1928                          * prepare_write() may have instantiated a few blocks
1929                          * outside i_size.  Trim these off again.
1930                          */
1931                         unlock_page(page);
1932                         page_cache_release(page);
1933                         if (pos + bytes > isize)
1934                                 vmtruncate(inode, isize);
1935                         break;
1936                 }
1937                 if (likely(nr_segs == 1))
1938                         copied = filemap_copy_from_user(page, offset,
1939                                                         buf, bytes);
1940                 else
1941                         copied = filemap_copy_from_user_iovec(page, offset,
1942                                                 cur_iov, iov_base, bytes);
1943                 flush_dcache_page(page);
1944                 status = a_ops->commit_write(file, page, offset, offset+bytes);
1945                 if (likely(copied > 0)) {
1946                         if (!status)
1947                                 status = copied;
1948
1949                         if (status >= 0) {
1950                                 written += status;
1951                                 count -= status;
1952                                 pos += status;
1953                                 buf += status;
1954                                 if (unlikely(nr_segs > 1)) {
1955                                         filemap_set_next_iovec(&cur_iov,
1956                                                         &iov_base, status);
1957                                         if (count)
1958                                                 buf = cur_iov->iov_base +
1959                                                         iov_base;
1960                                 } else {
1961                                         iov_base += status;
1962                                 }
1963                         }
1964                 }
1965                 if (unlikely(copied != bytes))
1966                         if (status >= 0)
1967                                 status = -EFAULT;
1968                 unlock_page(page);
1969                 mark_page_accessed(page);
1970                 page_cache_release(page);
1971                 if (status < 0)
1972                         break;
1973                 balance_dirty_pages_ratelimited(mapping);
1974                 cond_resched();
1975         } while (count);
1976         *ppos = pos;
1977
1978         if (cached_page)
1979                 page_cache_release(cached_page);
1980
1981         /*
1982          * For now, when the user asks for O_SYNC, we'll actually give O_DSYNC
1983          */
1984         if (likely(status >= 0)) {
1985                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
1986                         if (!a_ops->writepage || !is_sync_kiocb(iocb))
1987                                 status = generic_osync_inode(inode, mapping,
1988                                                 OSYNC_METADATA|OSYNC_DATA);
1989                 }
1990         }
1991
1992         /*
1993          * If we get here for O_DIRECT writes then we must have fallen through
1994          * to buffered writes (block instantiation inside i_size).  So we sync
1995          * the file data here, to try to honour O_DIRECT expectations.
1996          */
1997         if (unlikely(file->f_flags & O_DIRECT) && written)
1998                 status = filemap_write_and_wait(mapping);
1999
2000         pagevec_lru_add(&lru_pvec);
2001         return written ? written : status;
2002 }
2003 EXPORT_SYMBOL(generic_file_buffered_write);
2004
2005 ssize_t
2006 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2007                                 unsigned long nr_segs, loff_t *ppos)
2008 {
2009         struct file *file = iocb->ki_filp;
2010         struct address_space * mapping = file->f_mapping;
2011         size_t ocount;          /* original count */
2012         size_t count;           /* after file limit checks */
2013         struct inode    *inode = mapping->host;
2014         unsigned long   seg;
2015         loff_t          pos;
2016         ssize_t         written;
2017         ssize_t         err;
2018
2019         ocount = 0;
2020         for (seg = 0; seg < nr_segs; seg++) {
2021                 const struct iovec *iv = &iov[seg];
2022
2023                 /*
2024                  * If any segment has a negative length, or the cumulative
2025                  * length ever wraps negative then return -EINVAL.
2026                  */
2027                 ocount += iv->iov_len;
2028                 if (unlikely((ssize_t)(ocount|iv->iov_len) < 0))
2029                         return -EINVAL;
2030                 if (access_ok(VERIFY_READ, iv->iov_base, iv->iov_len))
2031                         continue;
2032                 if (seg == 0)
2033                         return -EFAULT;
2034                 nr_segs = seg;
2035                 ocount -= iv->iov_len;  /* This segment is no good */
2036                 break;
2037         }
2038
2039         count = ocount;
2040         pos = *ppos;
2041
2042         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2043
2044         /* We can write back this queue in page reclaim */
2045         current->backing_dev_info = mapping->backing_dev_info;
2046         written = 0;
2047
2048         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2049         if (err)
2050                 goto out;
2051
2052         if (count == 0)
2053                 goto out;
2054
2055         err = remove_suid(file->f_dentry);
2056         if (err)
2057                 goto out;
2058
2059         inode_update_time(inode, 1);
2060
2061         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2062         if (unlikely(file->f_flags & O_DIRECT)) {
2063                 written = generic_file_direct_write(iocb, iov,
2064                                 &nr_segs, pos, ppos, count, ocount);
2065                 if (written < 0 || written == count)
2066                         goto out;
2067                 /*
2068                  * direct-io write to a hole: fall through to buffered I/O
2069                  * for completing the rest of the request.
2070                  */
2071                 pos += written;
2072                 count -= written;
2073         }
2074
2075         written = generic_file_buffered_write(iocb, iov, nr_segs,
2076                         pos, ppos, count, written);
2077 out:
2078         current->backing_dev_info = NULL;
2079         return written ? written : err;
2080 }
2081 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2082
2083 ssize_t
2084 generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2085                                 unsigned long nr_segs, loff_t *ppos)
2086 {
2087         struct file *file = iocb->ki_filp;
2088         struct address_space *mapping = file->f_mapping;
2089         struct inode *inode = mapping->host;
2090         ssize_t ret;
2091         loff_t pos = *ppos;
2092
2093         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs, ppos);
2094
2095         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2096                 int err;
2097
2098                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2099                 if (err < 0)
2100                         ret = err;
2101         }
2102         return ret;
2103 }
2104
2105 ssize_t
2106 __generic_file_write_nolock(struct file *file, const struct iovec *iov,
2107                                 unsigned long nr_segs, loff_t *ppos)
2108 {
2109         struct kiocb kiocb;
2110         ssize_t ret;
2111
2112         init_sync_kiocb(&kiocb, file);
2113         ret = __generic_file_aio_write_nolock(&kiocb, iov, nr_segs, ppos);
2114         if (ret == -EIOCBQUEUED)
2115                 ret = wait_on_sync_kiocb(&kiocb);
2116         return ret;
2117 }
2118
2119 ssize_t
2120 generic_file_write_nolock(struct file *file, const struct iovec *iov,
2121                                 unsigned long nr_segs, loff_t *ppos)
2122 {
2123         struct kiocb kiocb;
2124         ssize_t ret;
2125
2126         init_sync_kiocb(&kiocb, file);
2127         ret = generic_file_aio_write_nolock(&kiocb, iov, nr_segs, ppos);
2128         if (-EIOCBQUEUED == ret)
2129                 ret = wait_on_sync_kiocb(&kiocb);
2130         return ret;
2131 }
2132 EXPORT_SYMBOL(generic_file_write_nolock);
2133
2134 ssize_t generic_file_aio_write(struct kiocb *iocb, const char __user *buf,
2135                                size_t count, loff_t pos)
2136 {
2137         struct file *file = iocb->ki_filp;
2138         struct address_space *mapping = file->f_mapping;
2139         struct inode *inode = mapping->host;
2140         ssize_t ret;
2141         struct iovec local_iov = { .iov_base = (void __user *)buf,
2142                                         .iov_len = count };
2143
2144         BUG_ON(iocb->ki_pos != pos);
2145
2146         down(&inode->i_sem);
2147         ret = __generic_file_aio_write_nolock(iocb, &local_iov, 1,
2148                                                 &iocb->ki_pos);
2149         up(&inode->i_sem);
2150
2151         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2152                 ssize_t err;
2153
2154                 err = sync_page_range(inode, mapping, pos, ret);
2155                 if (err < 0)
2156                         ret = err;
2157         }
2158         return ret;
2159 }
2160 EXPORT_SYMBOL(generic_file_aio_write);
2161
2162 ssize_t generic_file_write(struct file *file, const char __user *buf,
2163                            size_t count, loff_t *ppos)
2164 {
2165         struct address_space *mapping = file->f_mapping;
2166         struct inode *inode = mapping->host;
2167         ssize_t ret;
2168         struct iovec local_iov = { .iov_base = (void __user *)buf,
2169                                         .iov_len = count };
2170
2171         down(&inode->i_sem);
2172         ret = __generic_file_write_nolock(file, &local_iov, 1, ppos);
2173         up(&inode->i_sem);
2174
2175         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2176                 ssize_t err;
2177
2178                 err = sync_page_range(inode, mapping, *ppos - ret, ret);
2179                 if (err < 0)
2180                         ret = err;
2181         }
2182         return ret;
2183 }
2184 EXPORT_SYMBOL(generic_file_write);
2185
2186 ssize_t generic_file_readv(struct file *filp, const struct iovec *iov,
2187                         unsigned long nr_segs, loff_t *ppos)
2188 {
2189         struct kiocb kiocb;
2190         ssize_t ret;
2191
2192         init_sync_kiocb(&kiocb, filp);
2193         ret = __generic_file_aio_read(&kiocb, iov, nr_segs, ppos);
2194         if (-EIOCBQUEUED == ret)
2195                 ret = wait_on_sync_kiocb(&kiocb);
2196         return ret;
2197 }
2198 EXPORT_SYMBOL(generic_file_readv);
2199
2200 ssize_t generic_file_writev(struct file *file, const struct iovec *iov,
2201                         unsigned long nr_segs, loff_t *ppos)
2202 {
2203         struct address_space *mapping = file->f_mapping;
2204         struct inode *inode = mapping->host;
2205         ssize_t ret;
2206
2207         down(&inode->i_sem);
2208         ret = __generic_file_write_nolock(file, iov, nr_segs, ppos);
2209         up(&inode->i_sem);
2210
2211         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2212                 int err;
2213
2214                 err = sync_page_range(inode, mapping, *ppos - ret, ret);
2215                 if (err < 0)
2216                         ret = err;
2217         }
2218         return ret;
2219 }
2220 EXPORT_SYMBOL(generic_file_writev);
2221
2222 /*
2223  * Called under i_sem for writes to S_ISREG files.   Returns -EIO if something
2224  * went wrong during pagecache shootdown.
2225  */
2226 ssize_t
2227 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2228         loff_t offset, unsigned long nr_segs)
2229 {
2230         struct file *file = iocb->ki_filp;
2231         struct address_space *mapping = file->f_mapping;
2232         ssize_t retval;
2233         size_t write_len = 0;
2234
2235         /*
2236          * If it's a write, unmap all mmappings of the file up-front.  This
2237          * will cause any pte dirty bits to be propagated into the pageframes
2238          * for the subsequent filemap_write_and_wait().
2239          */
2240         if (rw == WRITE) {
2241                 write_len = iov_length(iov, nr_segs);
2242                 if (mapping_mapped(mapping))
2243                         unmap_mapping_range(mapping, offset, write_len, 0);
2244         }
2245
2246         retval = filemap_write_and_wait(mapping);
2247         if (retval == 0) {
2248                 retval = mapping->a_ops->direct_IO(rw, iocb, iov,
2249                                                 offset, nr_segs);
2250                 if (rw == WRITE && mapping->nrpages) {
2251                         pgoff_t end = (offset + write_len - 1)
2252                                                 >> PAGE_CACHE_SHIFT;
2253                         int err = invalidate_inode_pages2_range(mapping,
2254                                         offset >> PAGE_CACHE_SHIFT, end);
2255                         if (err)
2256                                 retval = err;
2257                 }
2258         }
2259         return retval;
2260 }
2261 EXPORT_SYMBOL_GPL(generic_file_direct_IO);